Mlreflect

Changed

- Updated all docstrings to that they can be read by the documentation generator Sphinx (using the Napoleon extension
for Google-style docstrings).

Added

- Added unit tests files `test_layers.py`, `test_data_generator.py`, `test_noise_generator.py` and
`test_preprocessing.py` as well as the test runner `runner.py`.
- Added functions to h5_tools module that can be used to save the noise and background generated with the `noise`
module to the save h5 file
- Added `NoiseGenerator` class to `training.noise_generator` which allows dynamic noise and background generation during
training.
- Added `InputPreprocessor` properties `has_saved_standardization`, `standard_mean`, `standard_std`.
- Added `naming` module with `make_timestamp()` function to create identifiers for training output.
- Added `check_gpu.py` script to be able to quickly check if tensorflow can find the GPU.

Changed

- Slightly changed the API of `OutputPreprocessor.restore_labels()`. It now only takes a single argument (the
normalized labels).
- Refactor internal package structure of the package source.
- Cleaned up a lot of attributes and methods and turned them into properties.
- Improved `has_saved_standardization` property of the `InputPreprocessor` class (return value depends now on whether
or not `standard_mean` and `standard_std` are `True` or `False`.)
- Improved `__repr__` of `Layer` and `MultilayerStructure` classes.
- Updated `usage_example.ipynb` to work with the new API.

Fixed

- Fixed several inconsistency bugs of the `label_removal_list` feature of the `OutputPreprocessor` class (with the
help of unit tests).
- Background and noise levels can now also be of `int` type (previously only `float`).

Added

- Added option for the `noise` module to generate shot noise and backgrounds with random levels within a given range.

Changed

- The thickness of the bottom most layer (substrate) is no longer a label, because it has no influence on the data
generation process and its presence was confusing.

Fixed

- Fixed bug where the roughness of the bottom most layer was dependent on its thickness (which in turn was always set
to 1). This led to a substrate roughness that was confined between 0 and 0.5 Å.

Added

- Added new C++-based reflectivity simulation engine from the refl1d package. This should be ~20 times faster than the
built-in code. The simulation engine can be chosen by using via the `engine` keyword of the ReflectivityGenerator class.
- Output normalization can now be changed from a [0, 1] range to a [-1, 1] range by choosing the approriate value for
the `normalization` keyword of the `OutputPreprocessor` class.
- Added method `InputPreprocessor.revert_standardization()`.

Changed

- The ambient SLD is now also given as a range (instead of a single value) and can be used as a non-constant label for
training.
- `OutputPreprocessor.apply_preprocessing()` now returns a tuple containing preprocessed labels in addition to the
removed labels both as pandas DataFrames.
- The `removed_labels` DataFrame is now used to by the `restore_labels` method instead of the previous `training_labels`
DataFrame (which caused some confusion).
- Removed methods and properties too reduce overhead and limit inconsistencies.
- Changed the .h5 file format that is with the `h5_tools` module.
- It is now no longer designed to save training, validation and testing data in separate groups to give the user
more flexibility. As a result, the group hierarchy was reduced by one level ("data" group was removed).
- All non-data information (units, min/max label values, etc.) have now been moved to the "info" group.
- Removed job list functionality from the `InputPreprocessor` class because it was unintuitive to use. Now the class is
only used for input standardization.
- Moved methods `apply_shot_noise()`, `generate_background()`, `apply_gaussian_convolution()` of the
`ReflectivityGenerator` class to a new `noise` module as stand-alone functions that can be applied to any previously
generated reflectivity curves.
- `apply_gaussian_convolution()` now uses the gaussian convolution from the refl1d package.
- Updated example notebook `data/notebooks/usage_example.ipynb` to match API of version 0.13.0

Fixed

- Non-constant labels that are removed via the `OutputPreprocessor` class are now not incorrectly added to the restored
labels anymore.
- Fixed that the wrong number of thin film layers was saved to the .h5 file when using the `h5_tools` module.
- The built-in reflectivity engine can now no longer generate intensities that are higher than 1.

Changed

- Changed all instances of keras to tensorflow.keras

Added

- Added CHANGELOG.md file to the project to track changes between releases
- Added method MultilayerStructure.rename_layer()
- Added docstrings to MultilayerStructure class

Changed
- SimpleModel now returns tuple (hist, timestamp) instead of only hist